Repository navigation
Conversation
Automated AI review
No defect found in the normal dispatcher path at Verified:
A stale ownership marker can restore 5 over another writer's 4If a dispatcher event arrives after the card was unbound or removed, the sysfs Reproduced with the production script and a stubbed Impact: a single write of the hybrid default while off 2.4 GHz, outside the 2.4 GHz case the issue describes. It contradicts the line 82 comment and the docs' "only if Suggested change: record the interface's Minor notes
Review informationTest scope: head AI process: Opus 5.5 Medium coordination and synthesis, Opus 5.5 Xhigh technical review and final fact check, GPT 6 Sol Xhigh search for related issues, Opus 5.5 Medium editorial check. Opt out: To stop receiving these reviews, reply to this comment saying so. |
|
Well yeah, "no defect found in the normal dispatcher path" still isn't a useful lead here. The question behind this patch is how the BCM4350 actually behaves with 2.4 GHz Wi-Fi and Bluetooth active. Code following the expected path doesn't establish that; the MacBookPro14,1 measurements below are the relevant evidence. The ownership finding was valid. I independently reproduced it, and the update fixes it. Calling the whole review "read only" was too broad, since you did run sandbox tests. I still think this could've been a much shorter, more useful review. For a hardware-specific patch, please test on the hardware where possible. For a sandbox bug report, include the runnable reproducer and output so the result can be checked directly. A long description of the testing isn't a substitute for sharing it. Lead with the actual defect and evidence, please. |
|
Here are the hardware measurements from the September 26 test on MacBookPro14,1, BCM4350 rev 5, firmware During continuous music playback, only
Expected cadence: 46.875 packets/s. The three mode-4 windows contain 1,207 packets, with no sequence gaps within any window. Playback reported Playing in 60/60 samples; 0/306 analyzed PCM blocks before Bluetooth encoding were silent. Listening matched the capture: steady music in mode 4, cutouts returning after restoring 5. The concurrent download transferred 8,861,121 bytes in 9.022 s, averaging 982,142 bytes/s. That establishes working Wi-Fi under this load; it does not measure parallel mode's throughput or latency cost against mode 5. Windows exclude the first 1 s and last 0.1 s of each phase. Completion delays use FIFO matching of host ACL submissions to controller completion events, accounting for eight initially outstanding packets; they are not end-to-end audio latency or measurements of RF packet loss. The 5→4→5 transitions were tested on hardware. The final PR entrypoint was separately exercised with mode 4 already active. These results support the firmware-policy change; they do not validate the separate stale-ownership/rebind case raised above. |
|
Thanks for the feedback. I agree that the way the review is structured was bad in this case and it has grown way too long. I've stopped the current run to improve it. |
|
Reviewed at What ran, on a disposable Omarchy VM (QEMU, not a Mac):
What was only read: the vendor command packing and reply parsing. Codex Medium, the second reviewer, checked them against brcmfmac's The evidence for the fix itself is your hardware comparison in #13297 and above. It has not been reproduced independently, and no worker can reproduce it. You opened the issue as well as this pull request, so the confirmation that would mark this verified needs to come from someone else: a maintainer, or another MacBookPro14,1 owner, testing this exact commit. One narrow case, left to you rather than pushed. At For the maintainer:
Related: #7644 also targets Broadcom Bluetooth audio dropouts on Macs, through the codec cap and btusb autosuspend. This machine's Bluetooth is on UART, and #7644 touches none of these files, so the two don't compete. Nothing else open changes Reviewed by Claude Opus 5.5 and Codex Medium. Codex's agreement on the packing and safety points was given with this review's reasoning available to it, so it isn't guaranteed to be independent. It found the narrow case above on its own. Waiting on the maintainer. |
AI-assisted-by: GPT-6.1 Sol high
|
Reproduced and fixed the ownership case. On
These tests use device/firmware fixtures, without live resets or policy changes. The existing 5→4→5 hardware comparison supports the policy; independent testing of this revision on another MacBookPro14,1 remains outstanding. Hardware/firmware guards and off-2.4 GHz restoration are unchanged. The kernel suggested in #13297 remains untested here; the author's September 28 question about its changes and intended comparison is unanswered. — GPT-6.1 Sol high |
|
| state.write_text(ifindex + '\n') | ||
| iovar(iface, 'btc_mode', 4) |
There was a problem hiding this comment.
Recovery loses override ownership
ifindex is checked before the firmware queries, but every iw command uses the interface name. If driver recovery replaces wlan0 during those queries, this branch can set the new interface to mode 4 while recording the old interface's index. The next event deletes that marker, leaving parallel mode active without permission to restore it on 5 GHz or disconnect.
Use the saved interface index for firmware commands, or stop when the interface changes. Add a test that replaces the interface during a query.
Fixes #13297.
On MacBookPro14,1, BCM4350's default hybrid coexistence policy interrupts A2DP audio while 2.4 GHz Wi-Fi is connected. Switching to parallel mode restored steady tones and music, including a concurrent Wi-Fi download; restoring hybrid mode reproduced the failure. The issue and hardware measurements document the comparison.
Add one NetworkManager dispatcher scoped to MacBookPro14,1, BCM4350 rev 5 (
14e4:43a3), and firmware7.35.180.133/01-c45b39d6. It changes default mode 5 to mode 4 on 2.4 GHz and restores only its owned override when disconnected or on another band. Ownership records the interface'sifindex, so a stale marker cannot authorize restoration after re-registration. A confirmed nonparallel readback after an attempted write releases ownership; an unavailable acknowledgement/readback retains recovery state because the outcome is unknown. Different firmware and nondefault policies are preserved.Repeated events with mode 4 active send no additional firmware write; irrelevant events exit before Python. A boot-only command would miss later firmware reloads, while leaving mode 4 enabled on every band broadens the tested policy. The existing
omarchy-settingsrecipe ships the executable directly frometc/, so package updates maintain or retire it and new/existing installs pick it up at their next network activation. No daemon, migration or driver reset.Validation:
786e1ffc. The full fixture sequence on that head confirms an ignored write leaves mode 5 and a marker, followed by an unwanted SET 5 after another tool chooses 4. With this correction the marker is absent and the independent mode 4 is preserved. No live network or firmware-policy writes were used for these tests.7b336b1b; those remain historical results, not a new aggregate pass.